DFS: A Feature-Sample Collaborative Optimization Framework for Machine Learning-Based Forest Aboveground Biomass...

Area of Science:

• Forestry
• Remote Sensing
• Ecology
• Geospatial Analysis

Background:

• Accurate forest aboveground biomass (AGB) estimation is vital for global carbon cycle monitoring and sustainable forest management.
• Existing machine learning methods face challenges with feature redundancy, uneven sample distribution, and inefficient hyperparameter tuning, limiting accuracy and efficiency.

Purpose of the Study:

• To develop a synergistic feature-sample optimization framework (DFS) for high-precision forest AGB estimation.
• To address limitations in feature selection, sample distribution, and hyperparameter optimization in current AGB estimation models.

Main Methods:

• Constructed Hunan and Hubei datasets using multi-source remote sensing and ground plot sampling.
• Developed Dual-Criteria Adaptive Feature Selection (DCAFS) using ReliefF and mutual information for optimal feature selection.
• Implemented Bidirectional Active Learning Sample Optimization (BALSO) for sample selection and spatial distribution optimization.
• Integrated Dream Optimization Algorithm (DOA) for adaptive hyperparameter tuning and data distribution alignment.

Main Results:

• The DFS framework achieved high accuracy on the Hunan (R2=0.83, RMSE=25.6 Mg·ha⁻¹) and Hubei (R2=0.86, RMSE=26.8 Mg·ha⁻¹) datasets.
• Validation on an independent dataset from Inner Mongolia confirmed the framework's effectiveness.
• The DCAFS method effectively reduced spectral redundancy while preserving biomass-sensitive information.

Conclusions:

• The DFS framework offers an effective and feasible approach for regional-scale forest AGB estimation.
• This method significantly improves estimation accuracy and computational efficiency compared to existing approaches.
• The study contributes to enhanced carbon monitoring and sustainable forest management strategies.